During one of our AI-fluency assessments, we could see exactly when AI entered her work.
It was when the report or presentation needed writing.
The conversation afterwards helped explain why. She was using AI exactly as she had been trained to.
She was not an AI sceptic. After a senior career at a multinational, she had founded a boutique consultancy working on ESG and cooling-systems optimisation. She already used AI for proposals, research summaries, presentations and client reports.
The assessment gave us a concrete piece of work to discuss. What we had observed there, and what she described afterwards, pointed to the same thing. The interesting part was not whether she used AI. It was where it entered the work.
In her day-to-day work, she would gather the site data, review the equipment information, do the analysis and reach a view herself. Then AI would help turn that view into a polished report. She would read the whole thing, check the facts and wording, and make the final call.
Her corporate training had taught her to use AI mainly to draft or improve content, treat every output as untrusted, and personally check the result before using it. Keep the human involved. Focus on the prompt. Use AI to save time on production.
That guidance made sense when organisations were first trying to introduce generative AI safely. But the method had become outdated. It kept AI at the writing stage and turned manual rereading into the main form of verification.
The method was still built around the draft
In the conversation after the assessment, she described what she normally checked when she reviewed a finished report. Mostly, she was checking the writing and whether the recommendation sounded right.
But in this kind of work, the awkward questions often sit further upstream. Are the operating hours complete? Are two energy figures using the same units? Is an emissions estimate using the right grid factor? Can the projected saving be recreated from the source data, or does it depend on an assumption that has disappeared into the spreadsheet?
A report can be beautifully written and still carry a weak baseline.
The assessment also showed where AI could have entered earlier. Its first job could be to question the data, flag missing readings or try to reproduce the savings calculation before she reviewed the recommendation.
That changes the frame. AI is no longer just helping to produce the deliverable. It can also help examine the evidence behind it: look for inconsistent units, separate measured values from estimates, or point out that a recommendation depends on weather or operating assumptions that have not been tested. The human still makes the decision, but no longer has to be the first and only line of review.
